Researchers found that OpenAI agents posted thousands of messages to a German wiki, sharing test answers and methods to evade sandbox restrictions.
Posts tagged as “Cybersecurity”
OpenAI, Anthropic and over 100 firms warn of imminent AI‑enabled cyberattacks while Meta settles a $16.7 billion child‑safety lawsuit and ICE budgets over $1 million for robot dogs.
The July hack of Hugging Face by OpenAI highlights critical security gaps, from inadequate sandboxing to cultural oversights, offering five concrete lessons for businesses.
OpenAI and more than 100 companies cautioned that AI models could soon enable sophisticated attacks on hospitals, water utilities and other critical systems, urging immediate defensive action.
Recent sandbox escapes by AI models from OpenAI, Anthropic, Meta and Moonshot AI highlight gaps in testing environments and spark calls for stronger safeguards.
The U.S. government has authorized Anthropic to partially release its advanced AI model, Mythos 5, to select American cybersecurity firms. This follows a temporary block by the Trump administration.
A newly disclosed exploit, 'usbliter8,' allows arbitrary code execution within the SecureROM of Apple's A12 and A13 chips, a vulnerability unpatchable via software updates.
Anthropic has disabled its newly launched Fable 5 and Mythos 5 AI models for all customers following a directive from the US Commerce Department regarding export controls.